Research map / Image generation / Fast diffusion sampling
Cross-Resolution Distribution Matching for Diffusion Distillation
Chen Feiyang, Hongpeng Pan, Haonan Xu, Xinyu Duan, Zhefeng Wang, Yang Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cross-resolution distribution matching (RM) to a cascaded 512->1024 student (ablation) | RMD on SDXL · HPS v2.1 prompts | HPS: 29.57 -> 31.99 | Improved |
| add RM at low-resolution target (no upsampling) (ablation) | RMD on SDXL · HPS v2.1 prompts (512px target) | HPS: 22.64 -> 30.28 | Improved |
| add RM to cascaded video generation (ablation) | RMD on Wan2.1-14B · VBench | VBench total: 81.11 -> 82.51 | Improved |
| add cascaded upsampling (UP) on top of RM (ablation) | RMD on SDXL · HPS v2.1 prompts | HPS: 30.28 -> 31.99 | Improved |
| use 2+2 step allocation instead of 3+1 (ablation) | RMD on PixArt-alpha · HPS v2.1 prompts | HPS: 24.79 -> 32.13 | Improved |
| use 1+3 instead of 2+2 allocation (ablation) | RMD on PixArt-alpha · HPS v2.1 prompts | HPS/speedup: 32.12 -> 32.13 HPS; 1.25x vs 1.68x speedup | No clear effect |
| mix predicted noise and Gaussian noise (alpha 0.2) in re-injection (ablation) | RMD on PixArt-alpha · HPS v2.1 prompts | generation fidelity | Improved |
| go from two to three resolution stages (256->512->1024) (ablation) | RMD on SDXL · HPS v2.1 prompts | HPS vs speedup: 31.99 -> 31.59 HPS; 1.56x -> 1.73x speedup | Mixed |
| go to four resolution stages (128->256->512->1024) (ablation) | RMD on SDXL · HPS v2.1 prompts | HPS: 31.99 -> 29.48 | Hurt |
| replace 4-step single-resolution TDM with 2+2 cascaded RMD | SDXL distillation · HPS v2.1 prompts | HPS average / speedup: 31.93 -> 31.99 HPS; 20.0x -> 33.4x | Improved |
| RMD vs 6-step distilled video baselines | Wan2.1-14B · VBench | Total score: 80.48 -> 82.51 | Improved |
| RMD vs the undistilled teacher on video | Wan2.1-14B · VBench | Total score: 83.75 -> 82.51 | Hurt |
Closest papers by meaning
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · similarity 0.90)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.89)
- Continuous-Time Distribution Matching for Few-Step Diffusion Distillation (NeurIPS 2026 · similarity 0.89)
- Improved Distribution Matching Distillation for Fast Image Synthesis (NeurIPS 2024 · similarity 0.89)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.88)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
